Skip to Main content Skip to Navigation

Apprentissage a contrario et architecture efficace pour la détection d'évènements visuels significatifs

Abstract : To ensure the robustness of a detection algorithm, it is important to get a close control of the false alarms it may produce. Because of the great variability of natural images, this task is very difficult in computer vision, and most methods have to rely on a priori chosen parameters. This limits the validity and applicability of the resulting algorithms. Recently, by searching for structures for which some properties are very unlikely to be due to chance, the a contrario statistical approach has proved successful to provide parameterless detection algorithms with a bounded expected number of false alarms. However, existing applications rely on a purely analytical framework that requires a big modeling effort, makes it difficult to use heterogeneous features and limits the use of data-driven search heuristics. In this thesis, we propose to overcome these restrictions by using statistical learning for quantities that cannot be computed analytically. The interest of this approach is demonstrated through three applications : segment detection, segmentation into homogeneous regions, and object matching from a database of pictures. For the two first ones, we show that robust decision thresholds can be learned from white noise images. For the last one, we show that only a few examples of natural images that do not contain the database objects are sufficient to learn accurate decision thresholds. Finally, we notice that the monotonicity of a contrario reasoning enables an incremental integration of partial data. This property leads us to propose an architecture for object detection which has an "anytime" behavior : it provides results all along its execution, the most salient first, and thus can be constrained to run in limited time.
Complete list of metadatas

Cited literature [118 references]  Display  Hide  Download
Contributor : Sophie Chouaf <>
Submitted on : Thursday, July 21, 2011 - 1:56:26 PM
Last modification on : Monday, October 19, 2020 - 11:13:02 AM
Long-term archiving on: : Sunday, December 4, 2016 - 8:38:00 PM


  • HAL Id : pastel-00610243, version 1


Nicolas Burrus. Apprentissage a contrario et architecture efficace pour la détection d'évènements visuels significatifs. Réseaux et télécommunications [cs.NI]. Université Pierre et Marie Curie - Paris VI, 2008. Français. ⟨pastel-00610243⟩



Record views


Files downloads